[juxtaposicion]

It is exciting that you could train a CLIP-style model from scratch with only 4M datapoints. But if you’ve got that data, why not fine tune a pretrained model with your 4M points? It seems likely to outperform the from-scratch method.

[vov_or]

There is not only a difference in the data source but pre-trained tasks as well. But you are right, a fine-tuned models on human-annotated data are way better than zero-shot (just pre-trained) on Image retrieval. And it is correct for CLIP, ALBEF, VICHA, and UFORM.

[ttt3ts]

Any plans to document how to fine tune your models then?

[vov_or]

It will take some time, but yes, we have this in our plans.

[riku_iki]

perhaps this approach can lead to better training of foundational models?..

[vov_or]

More efficient - for sure!